The activity of millions of neurons in the motor cortex drives the contraction of hundreds of muscles. Because there are so many more neurons than muscles, a vast number of different population activity patterns would likely generate the same movement. How does the brain choose from among these behaviorally equivalent options? To answer this question we analyze data from a brain-computer interface (BCI) center-out task where Rhesus macaques controlled a computer cursor using their neural activity, recorded in the primary motor cortex. In contrast to traditional approaches (e.g., monitoring neural activity while a monkey performs arm reaches), in a BCI the mapping from neural activity to cursor movement is defined by the experimenter. This allows one to define precisely the set of neural activities that will result in the same cursor movement. We find that the monkey's preferences for neural activity in this space vary with both task demands and behavioral output, despite the fact that activity patterns in this space all have the same effect on cursor movement. Our results suggest that the way redundancy is resolved at the level of neural population activity may be different than the way it is resolved at the level of the muscles.